Crunching The Numbers

As of 2018, we humans have collectively generated 33 trillion gigabytes of data, necessitating construction of enormous new data centers to store it and enormous power for machines to crunch it. Today, these centers alone emit as much as two percent of the world’s carbon dioxide.

But that’s just the start. According to one estimate, we’re on track to produce 175 trillion gigabytes of data by 2025; information may one day produce half the greenhouse gas emissions that the entire transportation sector currently does. Future computing centers may become so energy-intensive that just powering them up could actually limit computing performance.

That is, unless we replace energy-dissipating electrical components (where wires heat up via resistance) with frictionless optical components instead. “If you convert electrical wires to optical wires, then in principle, the energy dissipated in computing could be significantly lower,” says Michal Lipson, the Eugene Higgins Professor of Electrical Engineering.

Companies already use fiber optics to transfer data over long distances via light, but these bulky systems proved incompatible with data center technology. Thus Lipson, along with Charles Batchelor Professor of Electrical Engineering Keren Bergman, are developing components that route and interconnect networks of data encoded as light on a chip. For instance, Lipson recently developed an optical funnel that couples an optical fiber to a waveguide on a chip in order to transmit information between these components in a high-efficiency and high-bandwidth way, even if said components aren’t perfectly aligned. Lipson previously developed other influential components such as an all-optical switch in silicon—where light controls the flow of light in a circuit—and the slot waveguide, which can confine light to nanometer-wide regions of material with a low refractive index for use in these circuits.

Lipson’s team also works closely with Bergman’s group at the Lightwave Research Laboratory to construct photonic architectures for reducing power in data centers. This past fall, Bergman’s team won a $4.8 million, 3.5-year Defense Advanced Research Projects Agency grant to develop energy-efficient optical interconnects that can feed high-bandwidth signals communications from the chip to anywhere else in a computing system. They’re pushing forward this technology so photonic communications capabilities can integrate directly into data centers’ systems, such as their processors, GPUs, and memories. “Energy isn’t scalable,” Bergman notes, “and most of the energy is spent on moving data.”

For power hungry applications like AI, moving more communication infrastructure moves into the optical domain, could “open up vast capabilities in terms of both performance and the ability to sustain this sort of growth in data analytics and computation.” By avoiding the associated growth in energy consumption, photonic components may allow us to unleash the full potential of artificial intelligence without compromising our natural world.